BlockBeats News, March 3 — Research firm Quest Mobile released the "Core Report on AI Application Layer Development by 2025," which shows that by December 2025, the top five AI-native apps in the market by monthly active users are Doubao, DeepSeek, Yuanbao, Ant Afo, and Alibaba Qianwen. Ant Group's general AI assistant Lingguang, released in November last year, also entered the top ten. (Jin10)

The new paper released by the DeepSeek team introduces an inference system called DualPath, which optimizes large model inference performance, achieving a 1.87x increase in offline throughput and a 1.96x increase in online service. The application of large models is driving the evolution of forward intelligent agent systems, promoting a major transformation in human-computer interaction.

BlockBeats News, February 24 — According to market sources, AI startup DeepSeek is about to release a new AI model. CNBC analysts warn that if the model proves competitive and suppresses market sentiment, Nasdaq-listed stocks (especially AI-related stocks) could face a period of turbulence.

PANews January 27 News, according to Jintiao, DeepSeek has released the new DeepSeek-OCR 2 model, which adopts the innovative DeepEncoder V2 approach, enabling AI to dynamically rearrange different parts of an image based on its meaning, rather than simply scanning from left to right mechanically. This method simulates the logical process humans follow when viewing a scene. Ultimately, the model outperforms traditional vision-language models when processing images with complex layouts (such as documents or charts), achieving more intelligent and causally reasoning visual understanding.

Nvidia CEO Jensen Huang pointed out at the Davos Forum that in the past year, there have been three major breakthroughs in AI models: the emergence of agent-based AI that enhances reasoning capabilities, the release of open-source reasoning models like DeepSeek that promote ecosystem prosperity, and significant progress in physical AI in understanding language and physical phenomena.

DeepSeek, upon releasing the new model "MODEL1," also updated the FlashMLA code, involving 28 mentions of this model across 114 files, unlike V32. MODEL1 may be a new architecture, with main differences in memory optimization, KV cache layout, sparsity handling, and FP8 decoding.

BlockBeats News, January 14 — According to the latest data from Private Equity Data Platform, as reported by 21st Century Business Herald, DeepSeek founder Liang Wenfeng's Fantasia Quantitative's average return in 2025 reached 56.55%. It ranks second in the performance list of Chinese quantitative private equity firms with assets under management exceeding 10 billion yuan, only behind Lingjun Investment, which has an average return of 73.51%. Currently, Fantasia Quantitative manages over 70 billion yuan. According to Private Equity Data Platform, Fantasia Quantitative's average return over the past three years is 85.15%, and over the past five years is 114.35%. Some analysts believe that Fantasia Quantitative's impressive performance has provided sufficient R&D funding for Liang Wenfeng's DeepSeek. Fantasia Quantitative is one of the most well-known domestic quantitative private equity giants, founded by Liang Wenfeng in 2008 during his studies at Zhejiang University.

DeepSeek will release the new generation flagship AI model DeepSeek V4 in February, featuring excellent programming capabilities. Test results show that its performance surpasses top market models. The release is expected to take place around the Chinese New Year in 2025.

PANews January 1 News, according to Jin10 reports, DeepSeek has released a new paper proposing a novel architecture called Manifold-Constrained Hyperconnection (mHC), aimed at addressing issues such as training instability and limited scalability caused by the disruption of identity mapping properties in hyperconnection (HC) technology. The architecture restores the identity mapping characteristic by mapping the residual connection space of HC onto a specific manifold, while combining rigorous infrastructure optimization to ensure efficiency, achieving significant performance improvements and superior scalability. DeepSeek anticipates that mHC, as a flexible and practical extension of HC, will contribute to a deeper understanding of topological architecture design and point to promising directions for the evolution of foundational models. The paper is authored by Zhenda Xie, Yixuan Wei, Huanqi